There is Still Room to Improve Upon Epigenetic Clocks
Researchers here demonstrate that there is still room to improve the accuracy of epigenetic clocks based on patterns of DNA methylation. It could be argued that more effort should go towards generating a sufficient understanding of DNA methylation to allow the use of existing clocks to assess potential rejuvenation therapies, however. The long-term promise of epigenetic clocks is to provide a way to rapidly determine whether or not a given intervention is producing a meaningful reversal of aging, a replacement for time-consuming, expensive life span studies. Because researchers do not at present understand how underlying processes of aging are reflected in specific DNA methylation changes, it is impossible to take clock data at face value for any given intervention that targets only one aspect of aging. A clock may be insensitive to that mechanism of aging or it may give it too much weight. There is no way to know without undertaking life span studies to calibrate the clock against the intervention, defeating the point of the exercise.
"First generation" epigenetic ageing clocks, including those by Horvath and Hannum, were trained on chronological age (cAge), with near-perfect clocks expected to arise as sample sizes grow. However, cAge clocks hold limited capability for tracking and quantifying age-related health status, also termed biological age (bAge). To address this, "second generation" clocks have been trained on other age-related measures, including a phenotypic biomarker of morbidity (PhenoAge), rate of ageing (DunedinPACE), and time to all-cause mortality (GrimAge). Regressing an epigenetic clock predictor (whether trained on cAge or bAge) on chronological age within a cohort gives rise to an "age acceleration" residual with positive values corresponding to faster biological ageing.
Here, we sought to improve the prediction of both cAge and bAge. We first present large-scale epigenome-wide association studies (EWAS) of cAge (for both linear and quadratic CpG effects) and time to all-cause mortality as a proxy for bAge. A predictor of cAge is then generated using DNA methylation data from 11 cohorts, including samples from more than 18,000 participants of the Generation Scotland study. Through data linkage to death records in Generation Scotland, we develop a bAge predictor of time to all-cause mortality, which we compare against GrimAge, in four external cohorts. Our bAge predictor was found to slightly outperform GrimAge in terms of the strength of its association to survival. These analyses highlight the potential for large DNA methylation resources to generate increasingly accurate predictors of (i) cAge, with potential forensic utility, and (ii) bAge, with potential implications for risk prediction and clinical trials.